render/gles2: implement timer API

include/render/gles2.h

@@ -6,6 +6,7 @@
 #include <stdbool.h>
 #include <stdint.h>
 #include <string.h>
+#include <time.h>
 #include <wlr/render/egl.h>
 #include <wlr/render/gles2.h>
 #include <wlr/render/interface.h>
@@ -14,6 +15,11 @@
 #include <wlr/util/addon.h>
 #include <wlr/util/log.h>
 
+// mesa ships old GL headers that don't include this type, so for distros that use headers from
+// mesa we need to def it ourselves until they update.
+// https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/23144
+typedef void (GL_APIENTRYP PFNGLGETINTEGER64VEXTPROC) (GLenum pname, GLint64 *data);
+
 struct wlr_gles2_pixel_format {
 	uint32_t drm_format;
 	// optional field, if empty then internalformat = format
@@ -47,6 +53,7 @@ struct wlr_gles2_renderer {
 		bool EXT_texture_type_2_10_10_10_REV;
 		bool OES_texture_half_float_linear;
 		bool EXT_texture_norm16;
+		bool EXT_disjoint_timer_query;
 	} exts;
 
 	struct {
@@ -57,6 +64,12 @@ struct wlr_gles2_renderer {
 		PFNGLPUSHDEBUGGROUPKHRPROC glPushDebugGroupKHR;
 		PFNGLEGLIMAGETARGETRENDERBUFFERSTORAGEOESPROC glEGLImageTargetRenderbufferStorageOES;
 		PFNGLGETGRAPHICSRESETSTATUSKHRPROC glGetGraphicsResetStatusKHR;
+		PFNGLGENQUERIESEXTPROC glGenQueriesEXT;
+		PFNGLDELETEQUERIESEXTPROC glDeleteQueriesEXT;
+		PFNGLQUERYCOUNTEREXTPROC glQueryCounterEXT;
+		PFNGLGETQUERYOBJECTIVEXTPROC glGetQueryObjectivEXT;
+		PFNGLGETQUERYOBJECTUI64VEXTPROC glGetQueryObjectui64vEXT;
+		PFNGLGETINTEGER64VEXTPROC glGetInteger64vEXT;
 	} procs;
 
 	struct {
@@ -78,6 +91,15 @@ struct wlr_gles2_renderer {
 	uint32_t viewport_width, viewport_height;
 };
 
+struct wlr_gles2_render_timer {
+	struct wlr_render_timer base;
+	struct wlr_gles2_renderer *renderer;
+	struct timespec cpu_start;
+	struct timespec cpu_end;
+	GLuint id;
+	GLint64 gl_cpu_end;
+};
+
 struct wlr_gles2_buffer {
 	struct wlr_buffer *buffer;
 	struct wlr_gles2_renderer *renderer;
@@ -116,6 +138,7 @@ struct wlr_gles2_render_pass {
 	struct wlr_render_pass base;
 	struct wlr_gles2_buffer *buffer;
 	float projection_matrix[9];
+	struct wlr_gles2_render_timer *timer;
 };
 
 bool is_gles2_pixel_format_supported(const struct wlr_gles2_renderer *renderer,
@@ -128,6 +151,8 @@ const uint32_t *get_gles2_shm_formats(const struct wlr_gles2_renderer *renderer,
 
 struct wlr_gles2_renderer *gles2_get_renderer(
 	struct wlr_renderer *wlr_renderer);
+struct wlr_gles2_render_timer *gles2_get_render_timer(
+	struct wlr_render_timer *timer);
 struct wlr_gles2_texture *gles2_get_texture(
 	struct wlr_texture *wlr_texture);
 
@@ -140,6 +165,7 @@ void push_gles2_debug_(struct wlr_gles2_renderer *renderer,
 #define push_gles2_debug(renderer) push_gles2_debug_(renderer, _WLR_FILENAME, __func__)
 void pop_gles2_debug(struct wlr_gles2_renderer *renderer);
 
-struct wlr_gles2_render_pass *begin_gles2_buffer_pass(struct wlr_gles2_buffer *buffer);
+struct wlr_gles2_render_pass *begin_gles2_buffer_pass(struct wlr_gles2_buffer *buffer,
+	struct wlr_gles2_render_timer *timer);
 
 #endif

include/wlr/render/gles2.h

@@ -33,6 +33,7 @@ struct wlr_gles2_texture_attribs {
 };
 
 bool wlr_renderer_is_gles2(struct wlr_renderer *wlr_renderer);
+bool wlr_render_timer_is_gles2(struct wlr_render_timer *timer);
 bool wlr_texture_is_gles2(struct wlr_texture *texture);
 void wlr_gles2_texture_get_attribs(struct wlr_texture *texture,
 	struct wlr_gles2_texture_attribs *attribs);

render/gles2/pass.c

@@ -1,6 +1,8 @@
+#define _POSIX_C_SOURCE 199309L
 #include <stdlib.h>
 #include <assert.h>
 #include <pixman.h>
+#include <time.h>
 #include <wlr/types/wlr_matrix.h>
 #include "render/gles2.h"
 #include "types/wlr_matrix.h"
@@ -18,6 +20,19 @@ static struct wlr_gles2_render_pass *get_render_pass(struct wlr_render_pass *wlr
 static bool render_pass_submit(struct wlr_render_pass *wlr_pass) {
 	struct wlr_gles2_render_pass *pass = get_render_pass(wlr_pass);
 	struct wlr_gles2_renderer *renderer = pass->buffer->renderer;
+	struct wlr_gles2_render_timer *timer = pass->timer;
+
+	if (timer) {
+		// clear disjoint flag
+		GLint64 disjoint;
+		renderer->procs.glGetInteger64vEXT(GL_GPU_DISJOINT_EXT, &disjoint);
+		// set up the query
+		renderer->procs.glQueryCounterEXT(timer->id, GL_TIMESTAMP_EXT);
+		// get end-of-CPU-work time in GL time domain
+		renderer->procs.glGetInteger64vEXT(GL_TIMESTAMP_EXT, &timer->gl_cpu_end);
+		// get end-of-CPU-work time in CPU time domain
+		clock_gettime(CLOCK_MONOTONIC, &timer->cpu_end);
+	}
 	push_gles2_debug(renderer);
 	glFlush();
 	glBindFramebuffer(GL_FRAMEBUFFER, 0);
@@ -219,7 +234,8 @@ static const char *reset_status_str(GLenum status) {
 	}
 }
 
-struct wlr_gles2_render_pass *begin_gles2_buffer_pass(struct wlr_gles2_buffer *buffer) {
+struct wlr_gles2_render_pass *begin_gles2_buffer_pass(struct wlr_gles2_buffer *buffer,
+		struct wlr_gles2_render_timer *timer) {
 	struct wlr_gles2_renderer *renderer = buffer->renderer;
 	struct wlr_buffer *wlr_buffer = buffer->buffer;
 
@@ -240,6 +256,7 @@ struct wlr_gles2_render_pass *begin_gles2_buffer_pass(struct wlr_gles2_buffer *b
 	wlr_render_pass_init(&pass->base, &render_pass_impl);
 	wlr_buffer_lock(wlr_buffer);
 	pass->buffer = buffer;
+	pass->timer = timer;
 
 	matrix_projection(pass->projection_matrix, wlr_buffer->width, wlr_buffer->height,
 		WL_OUTPUT_TRANSFORM_FLIPPED_180);

render/gles2/renderer.c

@@ -1,3 +1,4 @@
+#define _POSIX_C_SOURCE 199309L
 #include <assert.h>
 #include <drm_fourcc.h>
 #include <GLES2/gl2.h>
@@ -5,6 +6,7 @@
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
+#include <time.h>
 #include <unistd.h>
 #include <wayland-server-protocol.h>
 #include <wayland-util.h>
@@ -18,6 +20,7 @@
 #include "render/gles2.h"
 #include "render/pixel_format.h"
 #include "types/wlr_matrix.h"
+#include "util/time.h"
 
 #include "common_vert_src.h"
 #include "quad_frag_src.h"
@@ -33,6 +36,7 @@ static const GLfloat verts[] = {
 };
 
 static const struct wlr_renderer_impl renderer_impl;
+static const struct wlr_render_timer_impl render_timer_impl;
 
 bool wlr_renderer_is_gles2(struct wlr_renderer *wlr_renderer) {
 	return wlr_renderer->impl == &renderer_impl;
@@ -52,6 +56,16 @@ static struct wlr_gles2_renderer *gles2_get_renderer_in_context(
 	return renderer;
 }
 
+bool wlr_render_timer_is_gles2(struct wlr_render_timer *timer) {
+	return timer->impl == &render_timer_impl;
+}
+
+struct wlr_gles2_render_timer *gles2_get_render_timer(struct wlr_render_timer *wlr_timer) {
+	assert(wlr_render_timer_is_gles2(wlr_timer));
+	struct wlr_gles2_render_timer *timer = wl_container_of(wlr_timer, timer, base);
+	return timer;
+}
+
 static void destroy_buffer(struct wlr_gles2_buffer *buffer) {
 	wl_list_remove(&buffer->link);
 	wlr_addon_finish(&buffer->addon);
@@ -543,18 +557,91 @@ static struct wlr_render_pass *gles2_begin_buffer_pass(struct wlr_renderer *wlr_
 		return NULL;
 	}
 
+	struct wlr_gles2_render_timer *timer = gles2_get_render_timer(options->timer);
+	clock_gettime(CLOCK_MONOTONIC, &timer->cpu_start);
+
 	struct wlr_gles2_buffer *buffer = get_or_create_buffer(renderer, wlr_buffer);
 	if (!buffer) {
 		return NULL;
 	}
 
-	struct wlr_gles2_render_pass *pass = begin_gles2_buffer_pass(buffer);
+	struct wlr_gles2_render_pass *pass = begin_gles2_buffer_pass(buffer, timer);
 	if (!pass) {
 		return NULL;
 	}
 	return &pass->base;
 }
 
+static struct wlr_render_timer *gles2_render_timer_create(struct wlr_renderer *wlr_renderer) {
+	struct wlr_gles2_renderer *renderer = gles2_get_renderer(wlr_renderer);
+	if (!renderer->exts.EXT_disjoint_timer_query) {
+		wlr_log(WLR_ERROR, "can't create timer, EXT_disjoint_timer_query not available");
+		return NULL;
+	}
+
+	struct wlr_gles2_render_timer *timer = calloc(1, sizeof(struct wlr_gles2_render_timer));
+	if (!timer) {
+		return NULL;
+	}
+	timer->base.impl = &render_timer_impl;
+	timer->renderer = renderer;
+
+	struct wlr_egl_context prev_ctx;
+	wlr_egl_save_context(&prev_ctx);
+	wlr_egl_make_current(renderer->egl);
+	renderer->procs.glGenQueriesEXT(1, &timer->id);
+	wlr_egl_restore_context(&prev_ctx);
+
+	return (struct wlr_render_timer *)timer;
+}
+
+static int gles2_get_render_time(struct wlr_render_timer *wlr_timer) {
+	struct wlr_gles2_render_timer *timer = gles2_get_render_timer(wlr_timer);
+	struct wlr_gles2_renderer *renderer = timer->renderer;
+
+	struct wlr_egl_context prev_ctx;
+	wlr_egl_save_context(&prev_ctx);
+	wlr_egl_make_current(renderer->egl);
+
+	GLint64 disjoint;
+	renderer->procs.glGetInteger64vEXT(GL_GPU_DISJOINT_EXT, &disjoint);
+	if (disjoint) {
+		wlr_log(WLR_ERROR, "a disjoint operation occurred and the render timer is invalid");
+		wlr_egl_restore_context(&prev_ctx);
+		return -1;
+	}
+
+	GLint available;
+	renderer->procs.glGetQueryObjectivEXT(timer->id,
+		GL_QUERY_RESULT_AVAILABLE_EXT, &available);
+	if (!available) {
+		wlr_log(WLR_ERROR, "timer was read too early, gpu isn't done!");
+		wlr_egl_restore_context(&prev_ctx);
+		return -1;
+	}
+
+	GLuint64 gl_render_end;
+	renderer->procs.glGetQueryObjectui64vEXT(timer->id, GL_QUERY_RESULT_EXT,
+		&gl_render_end);
+
+	int64_t cpu_nsec_total = timespec_to_nsec(&timer->cpu_end) - timespec_to_nsec(&timer->cpu_start);
+
+	wlr_egl_restore_context(&prev_ctx);
+	return gl_render_end - timer->gl_cpu_end + cpu_nsec_total;
+}
+
+static void gles2_render_timer_destroy(struct wlr_render_timer *wlr_timer) {
+	struct wlr_gles2_render_timer *timer = (struct wlr_gles2_render_timer *)wlr_timer;
+	struct wlr_gles2_renderer *renderer = timer->renderer;
+
+	struct wlr_egl_context prev_ctx;
+	wlr_egl_save_context(&prev_ctx);
+	wlr_egl_make_current(renderer->egl);
+	renderer->procs.glDeleteQueriesEXT(1, &timer->id);
+	wlr_egl_restore_context(&prev_ctx);
+	free(timer);
+}
+
 static const struct wlr_renderer_impl renderer_impl = {
 	.destroy = gles2_destroy,
 	.bind_buffer = gles2_bind_buffer,
@@ -573,6 +660,12 @@ static const struct wlr_renderer_impl renderer_impl = {
 	.get_render_buffer_caps = gles2_get_render_buffer_caps,
 	.texture_from_buffer = gles2_texture_from_buffer,
 	.begin_buffer_pass = gles2_begin_buffer_pass,
+	.render_timer_create = gles2_render_timer_create,
+};
+
+static const struct wlr_render_timer_impl render_timer_impl = {
+	.get_duration_ns = gles2_get_render_time,
+	.destroy = gles2_render_timer_destroy,
 };
 
 void push_gles2_debug_(struct wlr_gles2_renderer *renderer,
@@ -812,6 +905,16 @@ struct wlr_renderer *wlr_gles2_renderer_create(struct wlr_egl *egl) {
 		}
 	}
 
+	if (check_gl_ext(exts_str, "GL_EXT_disjoint_timer_query")) {
+		renderer->exts.EXT_disjoint_timer_query = true;
+		load_gl_proc(&renderer->procs.glGenQueriesEXT, "glGenQueriesEXT");
+		load_gl_proc(&renderer->procs.glDeleteQueriesEXT, "glDeleteQueriesEXT");
+		load_gl_proc(&renderer->procs.glQueryCounterEXT, "glQueryCounterEXT");
+		load_gl_proc(&renderer->procs.glGetQueryObjectivEXT, "glGetQueryObjectivEXT");
+		load_gl_proc(&renderer->procs.glGetQueryObjectui64vEXT, "glGetQueryObjectui64vEXT");
+		load_gl_proc(&renderer->procs.glGetInteger64vEXT, "glGetInteger64vEXT");
+	}
+
 	if (renderer->exts.KHR_debug) {
 		glEnable(GL_DEBUG_OUTPUT_KHR);
 		glEnable(GL_DEBUG_OUTPUT_SYNCHRONOUS_KHR);